Deep Knowledge Tracing Integrating Temporal Causal Inference and PINN

Knowledge tracing predicts students' future performance based on their historical performance data, which is significant for students' learning resource recommendation, learning path prediction, and other aspects. Students' knowledge mastery, learning ability, and question difficulty all influence the performance metrics of knowledge tracing. This paper proposes a deep knowledge tracing model that integrates temporal causal inference and the PINN (Physics-Informed Neural Network) model. The model first uses the temporal causality model to explores the causal relationships between students' knowledge points, which is then combined with the deep learning-based knowledge tracing model for prediction. Next, it treats the logical model as a 'physical model', adds a loss term, considers the confounding factors caused by students' answer preferences, and adjusts students' learning ability through backdoors to obtain more accurate predictions. In the public education datasets ASSISTment2012 and ASSISTchall, the predictive performance of the TLPKT-PINN model is superior to some classical models and LPKT. From the experimental results, we can conclude that considering the degree of mastery of causal knowledge points and adjusting the loss term can improve the accuracy of predicting student grades.